Users congratulated the PrimeIntellect team on Verifiers V1 for agentic RL because the decomposition improves eval frameworks and the launch marks a major technical milestone with strong compression gains.
Based on 20 visible X reactions from 33 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@PrimeIntellect congrats guys! especially @willccbb @xeophon
@willccbb @mikasenghaas @xeophon Congrats.
@PrimeIntellect amazing work!!
@PrimeIntellect Congrats!
this has been a labor of love for months with @mikasenghaas and @xeophon driving incredible progress, and `v1` is now finally ready for prime time we set out to fully modernize `verifiers` for the agent harness era, and unlocked some insane efficiency gains along the way https://x.com/willccbb/status/2076449433483616346/photo/1 https://twitter.com/PrimeIntellect/status/2076447247693402301
The rewritten stack splits environments into tasksets, harnesses, and runtimes.
@PrimeIntellect congrats guys! especially @willccbb @xeophon
@PrimeIntellect Congrats!
this has been a labor of love for months with @mikasenghaas and @xeophon driving incredible progress, and `v1` is now finally ready for prime time we set out to fully modernize `verifiers` for the agent harness era, and unlocked some insane efficiency gains along the way https://x.com/willccbb/status/2076449433483616346/photo/1 https://twitter.com/PrimeIntellect/status/2076447247693402301
well, and if we control the requests, we also have full control over the responses before they reach the harness 👀 i am currently playing with that concept, detecting + blocking requests that reward hack *while they happen*. this is what the interception server unlocks! https://x.com/xeophon/status/2076509933709733940/photo/1
but the interception server sends the actual request, not the harness. so you have full control. if you look at this screenshot, there is one tiny detail which codex simply doesn't expose: setting the temperature. well, verifiers v1 can do this. https://x.com/PrimeIntellect/status/2076447250763710692?s=20
@xeophon yeah that’s for sure useful. for our benchmarks/evals, we’ve been just playing an ad-hoc whack-a-mole game (which i think is fine for the current level of capabilities, but won’t be fine for much more capable agents).
@mikasenghaas @xeophon `uv add verifiers` read more: https://www.primeintellect.ai/blog/verifiers-v1
@xeophon ah that’s cool! but somehow i assumed you were planning to release a new hallucination benchmark 😅
full deep dive on verifiers v1: https://www.primeintellect.ai/blog/verifiers-v1
@willccbb @mikasenghaas @xeophon congrats!
Users congratulated the PrimeIntellect team on Verifiers V1 for agentic RL because the decomposition improves eval frameworks and the launch marks a major technical milestone with strong compression gains.
Based on 20 visible X reactions from 33 accounts; directional sample.
Ask a question below.
Published answers will appear here.
well, and if we control the requests, we also have full control over the responses before they reach the harness 👀 i am currently playing with that concept, detecting + blocking requests that reward hack *while they happen*. this is what the interception server unlocks! https://x.com/xeophon/status/2076509933709733940/photo/1
but the interception server sends the actual request, not the harness. so you have full control. if you look at this screenshot, there is one tiny detail which codex simply doesn't expose: setting the temperature. well, verifiers v1 can do this. https://x.com/PrimeIntellect/status/2076447250763710692?s=20
@xeophon yeah that’s for sure useful. for our benchmarks/evals, we’ve been just playing an ad-hoc whack-a-mole game (which i think is fine for the current level of capabilities, but won’t be fine for much more capable agents).
@mikasenghaas @xeophon `uv add verifiers` read more: https://www.primeintellect.ai/blog/verifiers-v1
@xeophon ah that’s cool! but somehow i assumed you were planning to release a new hallucination benchmark 😅